SpaceX (Nasdaq: SPCX) — After the First Print

The central structural insight

Of the three businesses SpaceX actually reports, exactly one earns durable rent — profit that survives the arrival of competitors and the cost of the capital tied up: Connectivity, where an orbital shell plus newly-acquired exclusive spectrum lets the company set its own price. The segment that produced the quarter's headline acceleration, AI, rests on cloud contracts either party may terminate on 90 days' notice — funded by capital commitments that cannot be cancelled at all.

Almost every published framing of this company — including the one that shaped market expectations into the print — still describes a launch company with a broadband arm and a rocket moonshot. Since 2 February 2026 it has been a three-segment enterprise — Space, Connectivity and AI, which in accounting terms are the units the chief executive actually allocates capital to and is judged on, not marketing labels — because on that date it absorbed X.AI Holdings Corp., which had itself absorbed X (formerly Twitter) eleven months earlier. The 10-Q — the quarterly report a US-listed company must file with the securities regulator — says so in its first substantive paragraph: the company "operate[s] three segments," the third of which "operates a vertically integrated AI platform spanning a frontier LLM Grok, AI solutions for consumer and enterprise customers, X … and AI computational infrastructure" (Form 10-Q, Note 1 — Nature of Business, filed 4 Aug 2026 (accession 0001628280-26-052535 index)). Any analysis that walks launch, Starlink and Starship and stops there is analysing roughly two-thirds of the revenue and one-seventh of the capital spending.

The capital budget says the same thing. In the June quarter the company spent $18,369 million on capital expenditure — cash laid out on long-lived assets such as satellites, rockets and servers, which lands on the balance sheet rather than the profit line. Of that, $15,828 million — 86.2% — went to AI. Launch infrastructure and the entire Starlink constellation together consumed 13.8% (Exhibit 99.1, Q2 2026 earnings release, segment capex table, 4 Aug 2026; filing index 0001628280-26-052515). Meanwhile the only segment that earned an operating profit was Connectivity, at $1,656 million on $4,291 million of revenue. Space lost $542 million. AI lost $1,257 million. Connectivity's profit did not merely carry the group; it funded $1,799 million of losses elsewhere and still left the consolidated line $143 million short of break-even.

So the market is being asked to pay $1.44 trillion — the value of every share at the current price — for a company in which one segment earns and two consume — and in which the consuming segment with the larger appetite has the weaker contracts. None of that dismisses the business. SpaceX has built things nobody else has built, and the Connectivity franchise is, on the evidence assembled below, one of the genuinely defended positions in modern infrastructure. It does demand precision about which part of the enterprise is an annuity — an income stream that keeps arriving without having to be won again each year — and which part is a bet.

Part ICompany & Domain Primer

Fifteen sections that build the business from first principles. A reader who finishes Part I should be able to check every claim in Part II without taking anything on trust.

01  What SpaceX now is — and why the segment map is the whole argument

Under US accounting rules, reportable segments follow how the chief operating decision maker actually allocates capital. They are not a presentation choice. SpaceX's 10-Q is explicit that the CEO is that person and that he manages the business as three units, judged on segment income from operations (Form 10-Q, Note 18 — Segments). So the segment map is a disclosure about internal reality, not marketing.

The three units, in the company's own words. Space "designs, manufactures, and launches reusable rockets"; its revenue comes only from third-party customers — launch services on Falcon 9 and Falcon Heavy, plus longer-cycle Launch & Development contracts for government programmes on Falcon, Starship and Dragon. Connectivity runs the Starlink broadband network and the Starshield government variant, selling to consumers, enterprises and states. AI runs Grok, the X platform, and — increasingly the point — a merchant compute business selling reserved GPU capacity to third parties (Form 10-Q, Note 1 and MD&A "Components of Results of Operations").

Space is a captive logistics arm with a merchant business attached, not the other way round. A merchant business sells to outsiders at whatever the market will bear; a captive one supplies its own parent, at whatever price the accountants set. The distinction decides where profit can even appear, because a captive arm has no price to earn a margin on. It has a cost. In the six months to June the company flew 78 launches, of which 17 were customer launches and 61 were internal, carrying 132 metric tons of customer payload against 908 tons of its own (Form 10-Q, MD&A "Key Business Metrics — Space"). Eighty-seven percent of the mass SpaceX put in orbit in the first half of 2026 was its own property.

One accounting sentence decides more of this analysis than any other disclosure in the set: "For launches of our Starlink satellites, the Company does not recognize any inter-segment revenue, rather those launch costs are capitalized in satellites in Property, plant, and equipment, net." Launch is transferred internally at cost. No markup, no notional market price, no transfer-pricingThe price one arm of a company charges another for goods or services supplied internally. Set it high and the profit shifts to the supplying arm; set it at cost, as SpaceX does, and the supplying arm shows none at all. The group total is unchanged either way — only the address of the profit moves. rent. Whatever surplus the launch business generates when it serves Starlink is invisible in the Space segment and shows up instead as a lower carrying value for satellites — which then depreciates through Connectivity's cost of revenue over the shell's life.

Capitalising a cost rather than expensing it is the mechanism doing the work here, and it is worth holding on to for the rest of the report. An expense hits this period's profit and is gone. A capitalised cost is parked on the balance sheet as an asset and then charged against profit in slices, over years, as depreciation. Nothing is hidden — the cash left the building either way — but it arrives in a different segment, in a different year, wearing a different name.

02  The road to the print: a merger, a split, a record IPO and an instant bond

Fourteen months compressed an unusual amount of corporate action into one balance sheet. On 28 March 2025, X.AI Holdings acquired X Holdings and X.AI Corp. On 2 February 2026, SpaceX acquired X.AI Holdings; both were accounted for as common-control mergersA combination of two businesses that were already owned by the same person or group. Because nothing genuinely changed hands, the accounts are restated as though the businesses had always been one — which is why last year's comparative figures already show a segment that did not exist inside the company last year., which is why the 2025 comparatives already contain an AI segment (Form 10-Q, Note 1 — Common Control Mergers). In May 2026 the board effected a five-for-one forward stock split. On 11 June the S-1The registration statement a company files with the SEC to sell shares to the public for the first time — the document that becomes the prospectus. "Going effective" is the regulator clearing it; only then may the shares be sold. went effective; on 12 June the shares began trading on the Nasdaq Global Select Market and Nasdaq Texas under SPCX.

The offering itself was 638,888,888 shares at $135.00, including full exercise of the underwriters' option, producing net proceeds of $85,675 million after $575 million of commissions and costs (Form 8-K, Item 8.01, 15 June 2026). The number circulating in the press — a "$75 billion IPO" — is the base deal: 555,555,555 shares × $135 = exactly $75.00bn, plus a 83,333,333-share greenshoeAn option letting the underwriters sell more shares than the announced deal — formally the over-allotment option — exercised when demand runs ahead of supply. Named after the first company to grant one. to reach $86.25bn gross. Underwriting economics on a deal this size collapse: $575 million on $86.25 billion is 0.67%, roughly a tenth of a conventional IPO gross spreadThe underwriters' fee, expressed as a percentage of the money raised. On an ordinary offering it is a meaningful slice of the proceeds; on a deal of this size the banks compete it down to a sliver, because a sliver of a deal that size is still a very good year..

Eleven days after listing the company came to the bond market for the first time, pricing $25 billion across five tranchesSlices of a single bond deal, each with its own size, interest rate and repayment date. Splitting an issue this way lets one borrower reach lenders who want their money back in five years and lenders who want it in thirty, in the same week. on 23 June: $7.0bn of 5.350% notes due 2031, $6.0bn of 5.650% due 2033, $6.0bn of 5.875% due 2036, $2.5bn of 6.600% due 2046 and $3.5bn of 6.650% due 2056 (Form 8-K, Item 8.01, 23 June 2026 (accession index)). Weighted average coupon 5.855%, weighted average maturity 11.7 years. The notes closed on 26 June under an indentureThe master contract governing a bond issue: what the borrower promises, what counts as a default, and what the lenders may do about it. The trustee is the party that enforces those terms on the bondholders' behalf, since thousands of them cannot each sue. with The Bank of New York Mellon Trust Company as trustee, alongside a registration rights agreement with the same five representatives, under which the company agreed to use commercially reasonable efforts to register the notes (Form 8-K, Items 1.01 and 2.03, 26 June 2026 (accession index)). The proceeds repaid a $20 billion bridge loanShort-term borrowing arranged to cover a gap until permanent financing arrives — here, until a newly public company could sell bonds in its own name. Bridges are expensive by design, because they are meant to be crossed quickly. taken in March to refinance the debt inherited with xAI — including a 12.5% secured note and two term loans that had cost the group $1,164 million in prepayment penalties ($425m + $221m + $518m) and produced a $1,526 million loss on extinguishmentThe accounting charge booked when debt is repaid before its due date. Penalties, fees and any costs not yet written off all land in one quarter, which is why clearing out expensive borrowing can make a period look far worse than the business underneath it. (Form 10-Q, Note 9 — Debt).

That single line explains most of the six-month loss. Net loss for the half was $4,817 million against $541 million for the June quarter alone; the gap sits almost entirely in "other income (expense), net" of $(1,962) million, driven by the extinguishment loss and an unrealised markdown on the company's 18,712 bitcoin, whose fair valueWhat an asset would fetch today if sold, rather than what was paid for it. Assets carried this way move the reported profit every quarter whether or not anything is bought or sold; the swing is "unrealised" until the asset actually changes hands. fell from $1,637 million at year-end to $1,098 million (Form 10-Q, Note 7 — Financial Instruments). The half-year headline is ugly for reasons that will not recur. The quarterly headline is the one to hold on to.

03  Launch economics from first principles — and the number that keeps getting misused

The physics of getting to orbit have not changed since Tsiolkovsky. The economics changed when the hardware stopped being thrown away. SpaceX's own framing in the prospectus is careful, and it is the source of a figure the industry repeats loosely: "according to NASA, the first version of Falcon 9 in 2010 had a launch cost of approximately $2,700 per kilogram, which represented a reduction of approximately 85% compared to the historical average launch cost per kilogram of $18,500. The first version of Falcon Heavy in 2018 further reduced this cost to approximately $1,400 per kilogram" (Final prospectus (424B4), "Business — Our Approach", 12 June 2026 (424B4 index)).

That sentence cites a NASA cost estimate for a 2010 vehicle and a 2018 vehicle. It is not a current price, not a current cost, and not a margin. The company does not disclose a current cost per kilogram, and no filing in this session's evidence base contains one. Every "SpaceX launches for $X per kilo today" claim in circulation is an estimate built on public list prices and assumed reuse counts, and should be treated as such.

The revenue side is another matter. In the June quarter, Launch Services revenue — recognised at a point in time, on delivery of the payload to orbit — was $648 million across 10 customer launches, or $64.8 million per customer launch, against $490 million across 9 launches a year earlier, or $54.4 million. Realisation per mission rose 19.1%, which management attributes to "a favorable customer mix shift" and larger payloads rather than to list-price increases (Form 10-Q, MD&A "Segment Results — Space"). That is mix, not pricing power, and the distinction is the whole point of this report.

04  The Falcon franchise today — and the cadence nobody flagged

Falcon launches in the June quarter were 37, down from 45 a year earlier — a decline of 17.8% that the coverage of the print almost entirely missed. Mass to orbit fell from 652 metric tons to 485, down 25.6%. Internal launches dropped from 36 to 27 (Form 10-Q, MD&A "Key Business Metrics — Space"). Customer launches went the other way, 9 to 10. So the fall is entirely in flights the company flew for itself.

The bearish reading is that the Falcon production and reuse system has hit a ceiling. The evidence is against it: the company launched 77 Falcons in the first half against 81 in the prior-year half, a 5% decline over six months, with the entire gap concentrated in the second quarter. The likelier reading — consistent with the disclosed pivot to V3 satellites launched on Starship — is that the company throttled Falcon-borne Starlink deployment while waiting for the vehicle that carries twenty times the downlink capacityHow much data the satellites carried on a single flight can push down to users on the ground. It is the real unit of Starlink production — the satellite is only the container, and what is being deployed is bandwidth. per flight. The prospectus states plainly that a single Starship launch "will be capable of deploying up to 60 V3 satellites to LEO, representing a potential twenty-fold increase in Starlink downlink capacity deployed relative to a Falcon 9 launch."

If that reading is right, the cadenceHow often a rocket family flies — flight rate, in the industry's own word. It is the number that matters most for a reusable fleet, because the enormous fixed cost of pads, factories and recovery ships is spread across whatever flies. dip is a deliberate pause before a step-change, and the first half of 2026 will be the low point. If it is wrong — if Starship slips again — then Starlink capacity growth slows precisely as three rival constellations are trying to close the gap. The September quarter's mass-to-orbit line will show which, and it costs nothing to check.

The fairest thing that can be said about the launch franchise sits in that last line. Space revenue of $962 million less cost of revenue of $329 million is $633 million — a 65.8% gross margin — on a cost base that was flat year over year while launch volumes fell 18%. Very few capital-goods businesses have that shape. The segment's reported loss is a choice: management is spending $1,076 million a quarter on Starship and running it through the same P&L line.

05  Starship — what has actually been demonstrated, and what has not

Starship is the vehicle on which three separate corporate futures depend: Starlink's V3 shell, NASA's return to the lunar surface, and — as Section 10 will show — a proposed orbital compute constellation of genuinely unhinged ambition. Cumulative investment disclosed in the prospectus is "over $15 billion." The programme now consumes more than a billion dollars a quarter and rising.

What has been demonstrated as of the evidence cutoff, from the company's own account: Flight 12 in May 2026 was the first suborbital mission of the V3 vehicle, achieving lift-off from the new Starbase pad, a precision landing of the upper stage, and deployment of modified V2 Starlink satellites. Flight 13, on 24 July 2026 — after the quarter closed — "achieved all flight objectives including deploying 20 production V3 satellites, demonstrating in-space relight of a Raptor engine, and executing the softest ever splashdown of Starship, providing critical views of an intact heatshield" (Exhibit 99.1, "Space Business Highlights", 4 Aug 2026).

Twenty V3 satellites, not sixty. A splashdown, not a catch. An intact heatshield observed, not a reflown vehicle. In-space relight demonstrated — which matters enormously, because relight is the precondition for the orbital manoeuvres that everything else needs. These are real milestones on a hard programme. They are not, on this evidence, "full and rapid reusability," which is the company's own stated destination and the thing the 99%-cost-reduction claim depends on.

Evidence: Exhibit 99.1 to the Form 8-K of 4 August 2026, "Space Business Highlights"; final prospectus, "Business" (V3 deployment capacity). Method: the right-hand column lists capabilities the company's own disclosures identify as necessary and which no source opened in this session records as achieved. Synthesis: the gap between the two columns is the Starship option's remaining execution risk, and it is where the R&D line is going.

The NASA relationship has become less comfortable. SpaceX won the Human Landing System award in 2021 to build a lunar Starship for Artemis III. On 20 October 2025, NASA's acting administrator said the agency would open the Artemis III lander contract to competition, stating: "SpaceX had the contract for Artemis 3. The problem is they're behind. They push their timelines out, and we're in a race against China" (SpaceNews, 20 Oct 2025). Blue Origin and Lockheed Martin were named as potential alternatives. The status of that re-competition as of the evidence cutoff could not be verified from a primary NASA source in this session and is treated accordingly below.

06  The Starlink shell — scale that is genuinely hard to replicate

As of 31 March 2026 SpaceX operated "approximately 9,600 Starlink broadband and mobile satellites in Low-Earth OrbitThe band of space a few hundred kilometres up, where a satellite circles the planet in roughly an hour and a half. Close enough that the signal round-trip feels instant, low enough that each satellite covers only a small patch of ground at a time — which is why the business needs thousands of them rather than a handful., which accounted for approximately 75% of all active maneuverable satellites (as of 31 Mar 2026) in orbit as of 31 March 2026," delivering median latency of about 25 milliseconds and median peak-hour residential download speeds of 225 Mbps across 164 countries, territories and markets (Final prospectus, "Business — Starlink Consumer Broadband"). A separate satellite-to-mobile texting and voice service reached approximately 7.4 million monthly unique devices across roughly 30 countries.

Three-quarters of everything manoeuvrable in orbit is a striking statistic, and it is not a broadband market share. It measures the physical and regulatory footprint one operator has established: orbital slotsThe orbital paths and frequencies a regulator has cleared an operator to use, coordinated internationally so two constellations neither collide nor jam each other. They are assigned by authorities, not bought from a previous owner, and a latecomer takes what is left. occupied, collision-avoidance obligations discharged, ground stations built, landing rightsA national government's permission to sell satellite service to customers inside its borders and to beam signals down to them. Every country grants its own, one at a time, on its own terms — and can withdraw it. negotiated in 164 jurisdictions. That footprint is the barrier, and it took roughly seven years and a captive rocket fleet to build. A rival with unlimited money still needs the launch capacity, the spectrum coordination and the market access, sequentially.

The nearest Western challenger is measurably behind, and a regulator has now put that in writing. Amazon's Kuiper Systems — rebranded Amazon Leo — was required under its FCC authorisation to deploy 1,616 satellites by 30 July 2026 and 3,232 by 30 July 2029. In its own waiver application it "estimated that it will have deployed approximately 700 satellites by July 30, 2026," which the Space Bureau observed is "approximately 21% of the … constellation" (FCC Order DA 26-553, adopted and released 5 June 2026, ¶¶ 28, 34). The Bureau granted a conditional waiver but demoted the processing-round spectrum priorityThe FCC licenses satellite systems in batches, and an earlier-approved system gets first call on a frequency: a later one must engineer around it, not the other way round. Losing that priority does not take your spectrum away — it makes you the party who has to give way. of every satellite launched after 30 July 2026 until the earlier of 30 March 2028 or the date Amazon reaches 1,616 in orbit.

The FCC's own language is unusually revealing about the regulator's incentive: "Amazon Leo is poised to offer cutting-edge satellite services worldwide as a new competitor-at-scale in low-Earth orbit," and the Bureau crafted "a remedy tailored to ensure that Americans quickly benefit from multiple, facilities-based providers." A regulator that grants a waiver in order to manufacture a second provider is telling you what it thinks of the first provider's market position.

07  Starlink unit economics — doubling subscribers while cutting price

The Connectivity segment is where the company earns, so its unit economics deserve arithmetic rather than adjectives. Subscribers reached 12.0 million at 30 June 2026, from 6.0 million a year earlier — up 101.2%, and up 1.7 million sequentially from 10.3 million at March. Over the same period average revenue per subscriber fell from $85 to $66 per month, a 22.4% decline that management attributes to "international expansion and the addition of lower priced service plans" (Form 10-Q, MD&A "Key Business Metrics — Connectivity" and "Segment Results — Connectivity").

Doubling volume while cutting price by a fifth is the classic signature of a business harvesting a cost advantage rather than exercising pricing power. Both can be excellent. They are not the same, and they age differently. A firm with pricing power raises price when demand grows; a lowest-cost producer cuts price to fill capacity it has already paid for and keeps the difference between its cost and the next-best producer's. Starlink is doing the second thing, and it is doing it deliberately — the marginal satellite is already in orbit and the marginal subscriber is nearly free.

One reconciliation shows how much of Connectivity is now not the consumer subscription business. ARPU of $66 across 11.15 million average subscribers implies about $2,208 million of subscriber service revenue in the quarter. Consumer revenue was $2,485 million, so roughly $277 million came from hardware and other consumer items. Enterprise & Government revenue — which includes aviation, maritime, land mobility, Starlink Mobile and Starshield, and which the company explicitly excludes from the subscriber count — was $1,806 million, or 42.1% of the segment and growing at 108% year over year. The fastest-growing part of the cash engine is the part with contracts, not the part with credit cards.

08  Starshield and the state — customer, regulator, subsidiser, and now a concentration risk

The 10-Q discloses two customers above 10% of consolidated revenue. Customer A was 18.3% of Q2 revenue and "relates to all three segments." Customer B was 19.5% and "relates to the AI segment" (Form 10-Q, Note 3 — Revenue, "Concentration of risk"). The filing does not name either. A customer buying launch, connectivity and AI services simultaneously, at roughly $1,430 million a quarter, is very likely the US government in aggregate — but the company does not say so, and this report treats the identification as an inference, not a fact.

The contract flow is disclosed. During the quarter SpaceX was "awarded over $6 billion in multi-year U.S. government contracts for Starshield … primarily from two major Space Force contracts for LEO-based communications and sensing constellations." Trade reporting identifies these as a $2.29 billion Other Transaction AuthorityA Pentagon contracting route that sits outside the ordinary federal acquisition rules, built to buy prototypes quickly from firms that will not submit to a conventional defence contract. Faster to award — and easier for either side to walk away from. agreement for the Space Data Network Backbone, requiring a fully operational prototype by end-2027 (SpaceNews, 26 May 2026), and a larger space-based moving-target-indicator award.

Those awards sit inside a larger programme label the market is beginning to use loosely. Golden Dome — the multi-layer homeland missile-defense architecture prioritised by the Department of Defense in 2025–26 — is the organising concept for a space-based sensing, tracking and communications layer, not a single line-item on SpaceX's revenue statement. Trade and defense-press reporting places SpaceX among the Other Transaction Authority awardees for space-based interceptor prototypes in a multi-vendor pool totaling up to $3.2 billion (DefenseScoop, 24 Apr 2026), and frames the May 2026 Space Force awards for a space-based moving-target indicator network and the $2.29 billion Space Data Network Backbone as the sensing and transport pieces of that layer (SpaceNews, 26 May 2026; contemporaneous trade accounts of a ~$4.16 billion SB-AMTI award). What is disclosed in SpaceX's own filings is the aggregate: "over $6 billion in multi-year U.S. government contracts for Starshield" in the quarter (Ex. 99.1, 4 Aug 2026). What is not disclosed is a Golden Dome revenue breakout. The analytical point does not need the brand name: government demand for proliferated LEO sensing and secure transport is real, multi-year, and concentrated — and it is also a two-edged dependence, because the same customer is regulator and, in spectrum and launch licensing, gatekeeper.

On launch, the national-security relationship is more revealing about pricing than about volume (award-level primaries live on SAM.gov and USASpending.gov; this section uses opened trade accounts of named awards rather than an unopened award PDF). For fiscal 2026 the Space Force assigned SpaceX five missions worth $714 million and ULA two worth $428 million, with Blue Origin receiving none pending certification of New Glenn; NSSLNational Security Space Launch — the Pentagon programme that buys launches for military and intelligence satellites. It certifies a small number of providers and then parcels missions among them in multi-year blocks, which is why an award schedule reveals more about relative standing than any press release. Phase 3 Lane 2 allocates roughly 28 of about 54 missions to SpaceX against 19 for ULA and up to 7 for Blue Origin, within a $13.7 billion FY2025–29 budget (SpaceNews, 4 Oct 2025).

The state is not only a customer. It is a regulator with the power to hand a competitor relief (Figure 6), a spectrum authority whose grants and denials move billions of enterprise value (Section 9), and a plaintiff-adjacent party: the United States has moved to intervene to dismiss an NAACP Clean Air Act suit seeking to enjoin the gas turbines powering the Colossus II data centre (Form 10-Q, Note 16 — Commitments and Contingencies, legal proceedings). A company this entangled with the state has a durable moat and a durable political exposure, and the same entanglement produces both.

09  Spectrum — the one asset a rival genuinely cannot manufacture

On 12 May 2026 the FCC's Wireless Telecommunications and Space Bureaus approved the assignment to SpaceX of approximately 65 megahertz of mid-band spectrumThe stretch of radio frequencies mobile networks prize above all others: low enough to travel a useful distance and get through walls, high enough to carry real data volume. Low-band reaches everywhere and carries little; high-band carries a great deal and barely crosses a street. Mid-band is the compromise everyone fights over. — AWS-4, AWS H-Block and unpaired AWS-3 licences — from EchoStar, through a two-step structure using Spectrum Business Trust 2025-1 (FCC Memorandum Opinion and Order DA 26-471, GN Docket No. 25-302, released 12 May 2026). The Commission described "the $17 billion transaction before us" as "merely one piece of evidence of the growing demand for this new type of connectivity."

SpaceX's own accounting puts the total higher: approximately $19.6 billion, comprising about $11.1 billion in equity — roughly 261.8 million Class A shares "at a fixed value of $42.40 per share" — and up to $8.5 billion for the payoff of designated EchoStar debt, plus separately disclosed Spectrum Credit Agreement payments of $1,241 million in 2026 and $828 million in 2027 that are structured as loans but "contemplated to be forgiven" and therefore booked as additional consideration (Form 10-Q, MD&A "Material Cash Commitments — Spectrum Transaction" and Note 6). $856 million had been paid as of 30 June. The two figures differ in scope, not in fact — the FCC's $17 billion excludes items SpaceX includes.

What SpaceX bought is the most defensible new asset on the balance sheet, and it is not yet on the balance sheet. The Commission granted waivers "of our service rules that will allow SpaceX to use the spectrum at issue flexibly for terrestrial or D2D services," and framed the outcome in competitive terms: "With a dedicated supply of D2D spectrum, we expect SpaceX to be able to compete with incumbent terrestrial wireless providers." This is the first time the company holds exclusive-use, contiguous, nationwide mid-band spectrum.

The grant came with teeth. SpaceX must hit interim performance requirements two years after 30 November 2027 — 300 kbps/MHz/beam downlink spectral efficiencyHow much data an operator pushes through each slice of airwave — kilobits per second, per megahertz of spectrum, per satellite beam. It measures engineering, not entitlement: two operators holding identical licences can deliver very different service from them., 100 uplink — and final requirements at nine years of 600 and 200 respectively, demonstrated per licence and per Basic Economic AreaOne of the geographic units the FCC uses to carve up the United States for licensing — roughly a city together with the territory that commutes to it. Buildout obligations are tested area by area, so hitting a national average is not a defence. with link budgets, coverage maps and customer measurements filed with the Commission. Miss the first interim and the final deadline accelerates by a year; miss the second and it accelerates by two more; miss the final and "SpaceX's license for such license area shall terminate automatically without Commission action, and SpaceX will be ineligible to regain it."

Two things follow. First, this is a genuine, transferable, exclusive property right of a kind no amount of capital creates — the FCC issues these, and it just gave SpaceX the only nationwide contiguous D2D block in the United States. Second, it is conditional, dated and forfeitable, which means it is an obligation to spend as much as it is an asset. The Commission was explicit that it imposed conditions "on par with a similarly situated terrestrial wireless network." SpaceX has effectively agreed to build a national mobile network from orbit, on a schedule, on pain of losing the licences.

The regulator has also shown it will say no. On 22 July 2026 the FCC declined to admit advanced satellite operations into the Upper C-band as part of the auction framework for 160 MHz of that spectrum, deferring SpaceX's request on the ground of transition complexity (SpaceNews, 1 July 2026, on the same meeting agenda; the specific C-band item is Tier 3 in this session and is not load-bearing). The same meeting adopted a Space Modernization Order replacing Part 25 with a new Part 100, compressing satellite application reviews — but pointedly, "pending applications for the massive orbital data center constellations proposed so far this year, such as SpaceX's plans for up to a million satellites … would not be subject to the order's revised processing timelines."

10  The AI segment — a merchant compute business wearing a consumer-internet coat

The AI segment contains three quite different things. X, the social platform, contributes advertising revenue that is shrinking: $367 million in the June quarter against $426 million a year earlier, down 13.8%, and $710 million for the half against $870 million, down 18.4% (Form 10-Q, Note 3 — Revenue disaggregation). Grok subscriptions and API access contribute a modest, undisclosed slice. And then there is the thing that actually moved the quarter.

"AI Solutions & Infrastructure" revenue was $2,194 million, up from $311 million a year earlier and $475 million in the March quarter. The earnings release attributes $1.6 billion of the increase to newly signed Cloud Services AgreementsContracts under which a customer reserves a defined block of computing capacity — named machines, a fixed monthly fee, a stated period — rather than paying for whatever it happens to use. The customer buys certainty of supply; the supplier takes the risk of having bought the hardware.. Nameplate compute draw — GPUs installed multiplied by their all-in power draw — reached 1.4 GW, from 1.0 GW at March and 0.4 GW a year earlier (Form 10-Q, MD&A "Key Business Metrics — AI").

The balance sheet tells the story more bluntly than the income statement. Gross property, plant and equipment stood at $83,071 million at 30 June. "Servers and networking equipment" was $34,771 million of it. "Satellites" was $13,788 million.

And the ambition runs further than terrestrial data centres. On 30 January 2026 the company filed with the FCC's Space Bureau for authority to operate an orbital data-centre constellation of up to one million satellites, each carrying roughly 100 kW of compute — a filing the Commission accepted in February and which, as noted above, is explicitly excluded from the new streamlined processing timelines. Independent verification of the application's specific parameters from an FCC primary was not obtained in this session; the exclusion itself is confirmed by the SpaceNews account of the 22 July order. Whatever one makes of the plan, it clarifies the corporate logic: Starship exists to move mass cheaply, and the heaviest thing anyone now wants to move is compute.

11  The compute contracts, as written — the most important paragraphs in the filing set

Two agreements produced the quarter's acceleration, and both are disclosed in enough detail to be assessed as legal instruments rather than as press releases.

Anthropic. Signed 3 May 2026. Access to compute capacity across Colossus and Colossus II comprising "approximately 325,000 NVIDIA GPUs, backed by hyperscale-class CPUs, exabyte-scale storage and high-speed networking." The customer "has agreed to pay us $1.25 billion per month through May 2029, with capacity ramping in May and June 2026 at a reduced fee." And then: "After the initial three-month period, the agreements may be terminated by either party upon 90 days' notice." (Final prospectus, "Business — Compute Services Agreements with Third Parties", and notes to the interim financial statements, "Cloud Services Agreement".)

Google. Signed 5 June 2026 and disclosed by free writing prospectusA short document a company mid-offering may file with the SEC to put new information in front of investors between the main prospectus and the sale. It is a supplement rather than a summary — which is why a contract of this size surfaces in one at all. the same day. "Approximately 110,000 NVIDIA GPUs, CPUs, memory, and other related components." Google "has agreed to pay us $920 million per month from October 2026 through June 2029 (SpaceX FWP; TechCrunch, 5 Jun 2026), with capacity ramping up through September at a reduced fee." If SpaceX fails to deliver the committed GPUs by 30 September 2026, then after a one-month grace period "Google may immediately terminate the agreement or accept the number of GPUs provided, with a corresponding pro rata reduction in the monthly fees." And: "After December 31, 2026, the agreement may be terminated by either party upon 90 days' notice." (Free Writing Prospectus filed pursuant to Rule 433, 5 June 2026.)

Group backlog at 30 June was $47,461 million, of which $14,286 million already sits as deferred revenueMoney a customer has already handed over for something not yet delivered. It sits on the balance sheet as a liability — an obligation to perform — and only becomes revenue as the service is actually supplied., with 56% expected to convert within a year (Form 10-Q, Note 3 — Backlog). Total backlog across all three segments is less than two-thirds of the nominal value of the two compute contracts alone. That is not an inconsistency; it is the accounting working correctly. Backlog counts what a customer is legally obliged to pay. On a 90-day notice period, that is one quarter.

A third arrangement is stranger still. On 19 April 2026 SpaceX entered a compute agreement with Anysphere, Inc. (Cursor) under which SpaceX provides GPU capacity and Cursor contributes "personnel, data and datasets, documentation, technical know-how, workflows, prompts, specifications and software code," with joint development of models. Concurrently it took a call option to acquire Cursor at an implied equity value of $60.0 billion. If SpaceX terminates the option, Cursor is owed a $1.5 billion termination fee and an $8.5 billion deferred services fee. The option was exercised and a merger agreement signed on 16 June 2026 (Form 8-K, Item 1.01, 16 June 2026). A $10 billion break-fee on an option that the 10-Q records as having "an initial fair value of zero" is an unusual instrument, and its consequence for dilution is quantified in Part II.

12  Financing the build — the bond, the obligations, and the leases nobody is discussing

SpaceX ended the quarter with $100,009 million of cash and marketable securities, $38,433 million of debt principal and an undrawn $5,000 million revolverA committed credit line a company may draw on and repay at will, like a corporate overdraft. Left undrawn it costs a small standby fee and functions as insurance; drawn, it is simply debt., and told investors it has "sufficient sources of funding to meet our business requirements for at least the next twelve months" (Form 10-Q, MD&A "Liquidity and Capital Resources"). On the face of it, fortress liquidity.

The commitments run the other way. Non-cancellable contractual obligations "primarily related to the Company's investments in AI infrastructure and third-party cloud capacity arrangements" total $27,955 million, of which $22,244 million — 79.6% — falls due in 2027 (Form 10-Q, Note 16 — Commitments and Contingencies, unconditional obligations). Add the spectrum obligations: $828 million of credit-agreement payments in 2027, up to $8.5 billion of EchoStar debt payoff at the Spectrum Acquisition Closing, and a further $827 million if closing slips to November 2028.

So the asymmetry that defines this balance sheet is not leverage. It is duration mismatch of obligation versus entitlement. Duration here means nothing more than how far into the future a promise binds. A business is safe when what it is owed lasts at least as long as what it owes; reverse the two and every quarter's income has to be won again while the bills arrive on a schedule set years ago. The revenue is terminable in 90 days. The spending is contracted for two years.

Then there are the related-party leases, which are the least-discussed disclosure in the document. In April 2026 CTC Property, LLC — a SpaceX subsidiary — entered an equipment lease with Valor Equity Partners for AI infrastructure hardware. Antonio J. Gracias, Valor's founder, CEO and chief investment officer, sits on the SpaceX board. The transaction was "deemed to be a failed sale-leaseback," meaning the assets stay on SpaceX's balance sheet and the payments are recorded as debt. As of 30 June, three such transactions carried $2,039 million of current and $11,290 million of non-current debt, with $327 million of interest expense in the quarter and $513 million for the half (Form 10-Q, Note 17 — Related Party Transactions). The prospectus discloses that the underlying Valor leases provide for aggregate cash payments of $6,986 million, $6,633 million and a further amount across the individual agreements.

$327 million of quarterly finance charge against a period-end balance of $13,329 million annualises to at least 9.8%. Because that balance grew sharply during the quarter — it was $4,507 million at December — the average-balance calculation gives roughly 14.7%. The true implied rate lies between, and it is not a pure borrowing rate: a failed sale-leaseback charge blends financing cost with an equipment-return component. But the comparison is still stark. The company borrowed $25 billion in the public market at a 5.855% weighted-average coupon while holding $100 billion in cash, and is simultaneously financing AI hardware through a director-affiliated vehicle at an implied all-in charge several points wider.

13  Control and capital structure — a founder with 84.4% of the votes and no sunset

At 28 July 2026 there were 7,696,293,669 Class A shares and 5,485,486,276 Class B shares outstanding — 13,181,779,945 in total (Form 10-Q, cover page). Class A carries one vote; Class B carries ten. On the 30 June balances that is 55,690 million Class B votes against 7,607 million Class A — 88.0% of the voting power in the high-vote class.

The prospectus beneficial-ownershipWho actually controls a shareholding and the votes attached to it, regardless of whose name appears on the certificate. Shares held through trusts, funds and nominees are counted to the person who directs them, which is why one line in the table can gather up dozens of separate vehicles. table lists exactly one holder above 5%: Elon Musk, with 849,494,440 Class A shares and 5,569,053,075 Class B shares, giving 84.4% of combined voting power after the offering. All executive officers and directors as a group hold 85.3%. The Musk holding includes 1,302,072,285 restricted Class B shares subject to performance conditions, 350,000,000 Class B shares issuable on options, and shares held through the Elon Musk Revocable Trust and several other trusts (Final prospectus, "Security Ownership of Certain Beneficial Owners and Management", note 1). The Form 3The filing an insider — an officer, a director, or anyone holding more than ten percent — makes on first becoming one, declaring what they own. Every later change is reported on a Form 4, which is how the market watches insiders buy and sell. filed on 11 June confirms the officer/director/ten-percent-owner status (Form 3, Elon Musk, event date 11 June 2026).

The charter is unusually tight. Class B converts automatically to Class A on any transfer that is not a "Permitted Transfer," and the corporation "shall not issue or sell any shares of Class B Common Stock … to any person that is not the Founder or his Permitted Transferee" (Restated Certificate of Formation, Exhibit 3.1 to the Form 10-Q, §§ 1, 6, 7, 13). No time-based sunset provision appears in the certificate. Control does not decay with time; it decays only if the founder sells, and Section 14 shows he cannot until June 2027.

The second-largest disclosed position belongs to entities affiliated with a director. The Valor Entities — 27 separate funds and vehicles enumerated in the prospectus footnote — hold 503,414,530 Class A shares, 6.7% post-offering, beneficially attributed to Antonio J. Gracias (Form 3, Antonio J. Gracias, 11 June 2026). The same director's firm is the counterparty to $13.3 billion of the company's equipment financing. Both facts are properly disclosed; together they describe an unusually concentrated web.

The board's posture shows in its own pay: non-employee directors "do not currently receive cash or equity compensation for their service on the Board or its committees" (Form 8-K, Item 5.02, 17 June 2026). The same filing announced Roelof Botha of Sequoia Capital joining the board and the audit committee on 16 June, disclosing that a Botha family member has worked at the company since January 2025 with compensation above the $120,000 reporting threshold. Directors Donald Harrison (Google), Steve Jurvetson and Roelof Botha reported no beneficial ownership on their Forms 3.

14  The lock-up as designed — a staircase, not a cliff

The lock-up architecture is the most carefully engineered part of this offering, and it is fully disclosed. Why a lock-up exists at all is worth a moment. A newly listed stock has almost no shares in circulation, and letting the pre-IPO owners out at once would drown the price before the market has worked out what the thing is worth; so the underwriters extract a promise not to sell. The design of that promise, far more than its length, decides how the stock behaves for its first year. Rather than a single 180-day cliff, the underwriters built two pools with automatic timed releases and one price-contingent tranche. Part II analyses what it means; this section establishes what it says.

Pool one — the 180-day lock-up. All shares not otherwise committed are restricted until the close of the 180th day after the prospectus date (8 December 2026), subject to automatic early releases: 20% on the second full trading day after the first earnings release; a conditional further 10% on the same date if a price test is met; 7% on each of 20 August, 9 September, 24 September, 9 October and 24 October 2026; 28% on the second full trading day after the Q3 results; and the balance on 8 December (Final prospectus, "Underwriting — Lock-up Agreements").

Pool two — the extended lock-up. Certain shareholders agreed to hold until the second full trading day after the June 2027 quarterly release, with automatic releases of 20% after the Q4 2026 results, 10% on 18 March 2027, 20% after the Q1 2027 results, 10% on 17 May 2027, 20% on 12 June 2027, and the remainder after the Q2 2027 results.

The founder. Musk agreed with the underwriters to a flat 366-day lock-up expiring 12 June 2027. The prospectus is explicit: "Our founder's shares are not subject to any early release provisions during the lock-up period." His release covers up to 6.4 billion Class A-equivalent shares, including 350 million underlying options.

The two long-dated pools together "represent approximately 7.8 billion shares of common stock, or greater than 63% of our shares outstanding immediately prior to this offering" — described elsewhere in the same document as "approximately 60% of our shares outstanding (after giving effect to this offering)." Both figures are correct; they use different denominators, and the difference is the 638.9 million new shares. The arithmetic reconciles: Musk's 6.4 billion less 350 million option shares, plus the roughly 1.76 billion extended-pool shares implied by the 20%/351.9 million tranches, gives 7.81 billion.

Finally, the disclosed intentions. On 12 June 2026 — the day of listing — the Valor Entities adopted a Rule 10b5-1A written plan that sets future share sales in advance — how many, at what price, on what dates — adopted at a moment when the insider holds no inside information. Once running, the sales execute automatically; that automation is the insider's defence against a later accusation of trading on what they knew. arrangement "for the potential distribution to limited partners and general partners of the Valor Entities of up to 225,857,490 shares of Class A common stock," expiring 30 September 2027. The CFO adopted a plan on 16 June covering up to 919,497 shares that "does not commence sales until 2027"; the President and COO adopted one on 23 June covering up to 585,605 shares (Form 10-Q, Item 5 — Other Information, Rule 10b5-1 Trading Arrangements). Both officers "agreed to subject the vast majority of [their] shares to the extended lock-up period."

On named venture and strategic holders the filings are more silent than the market's folklore. The prospectus beneficial-ownership table lists exactly one >5% holder: Elon Musk; the only other concentrated position the documents make concrete is the Valor Entities block attributed to director Antonio J. Gracias (503.4 million Class A shares, 6.7% post-offering) (Final prospectus, "Security Ownership of Certain Beneficial Owners and Management"; Form 3, Antonio J. Gracias, event 11 Jun 2026). Brands often recited in secondary commentary — Founders Fund, Sequoia, a16z, Google, Fidelity and peers — do not appear as >5% beneficial owners in the opened primary tables. They may hold below the disclosure threshold, through vehicles, or may have sold pre-IPO; this report does not invent a holder list the S-1 does not print. Musk's post-IPO Form 4 records the preferred conversions and the continuing 350 million Class B options at a $8.3998 exercise price, not open-market distribution (Form 4, Elon Musk, filed 17 Jun 2026). Early-release authority sits with Goldman Sachs & Co. LLC on behalf of the underwriters, subject to the prospectus exceptions — a contractual residual, not a scheduled tranche.

15  The competitive field, checked against each rival's own filings

Sector reasoning is cheap. Each named competitor's own primary disclosure says something more specific.

Amazon (Amazon Leo, formerly Project Kuiper (launch-schedule compilation, Orbital Radar — Tier 3 for cadence tallies only)). Amazon's own 10-Q discloses that prepaid expenses "include amounts related to satellite network launch services deposits" of $7.4 billion at 30 June 2026, expensed on launch to "Technology and infrastructure" (Amazon.com, Inc. Form 10-Q for the quarter ended 30 June 2026, Note 1; on the deployment clock, Via Satellite / Satellite Today, 5 Jun 2026). Amazon does not report Leo revenue or subscribers separately. More consequentially, on 13 April 2026 Amazon signed a definitive merger agreement to acquire Globalstar at an implied value of approximately $10.9 billion including debt, with agreements alongside it with Apple, Globalstar's largest customer, expected to close in 2027 (Amazon Form 10-Q, "Other Contingencies" and acquisition disclosure). Amazon is buying mobile-satellite spectrum and the Apple relationship in the same season SpaceX bought EchoStar's. That is the clearest possible confirmation of what both companies think the scarce asset is.

EchoStar. The counterparty to the spectrum deal is under stress. On 2 August 2026 — two days before SpaceX filed its 10-Q — EchoStar's subsidiary Hughes Satellite Systems Corporation and eleven affiliates filed voluntary Chapter 11The US bankruptcy chapter under which a company keeps trading while it renegotiates its debts under court supervision, rather than being broken up and sold off. Filing freezes creditors where they stand and puts a judge between the company and its lenders. petitions in the US Bankruptcy Court for the District of Texas, triggering automatic acceleration of HSSC's 5.25% secured and 6.625% senior notes (EchoStar Corporation and Hughes Satellite Systems Corporation Form 8-K, Items 1.03, 2.04 and 8.01, filed 3 August 2026). EchoStar Corporation itself did not file. SpaceX's 10-Q contains no reference to the filing anywhere in its spectrum disclosure. Part II assesses whether it should.

Eutelsat / OneWeb. OneWeb operates 648 satellites. Trade reporting places Eutelsat's LEO connectivity revenue at roughly €280 million for the year to 30 June 2026, growing about 50%. A Eutelsat primary filing was not opened in this session; the figure is Tier 3 and is used only to establish an order of magnitude, and no verdict in this report rests on it.

AST SpaceMobile. The pure-play direct-to-device challenger reported total revenues of $14.7 million for the three months ended 31 March 2026, against $0.7 million a year earlier, holding 38 patent families and roughly 3,900 patent claims worldwide (AST SpaceMobile, Inc. Form 10-Q for the quarter ended 31 March 2026). Its intellectual-property position is real; its revenue is a rounding error against SpaceX's Connectivity segment. Note the period mismatch — AST's quarter ends three months before SpaceX's.

Rocket Lab. Total revenue of $200.3 million for the quarter ended 31 March 2026 against $122.6 million a year earlier, with company-wide remaining backlog of $2,219.8 million, 36% expected within twelve months, and the reusable medium-lift Neutron vehicle still in development (Rocket Lab Corporation Form 10-Q for the quarter ended 31 March 2026, Note 3 and Item 2). The comparison to SpaceX's $47,461 million backlog is not like-for-like — Rocket Lab's is a launch-and-space-systems company; SpaceX's spans three segments including compute.

Tesla. Not a competitor but a related party on both sides. Tesla invested $2.00 billion in SpaceX common stock in March 2026 (converted from an xAI preferred position), holds "less than 1%," and marked the stake to $3,007 million at 30 June under a fair-value election, booking a $1.00 billion gain in the quarter; it is subject to IPO-related sale restrictions expiring December 2026. Tesla separately recognised $318 million of Megapack revenue from SpaceX in the quarter (Tesla, Inc. Form 10-Q for the quarter ended 30 June 2026, Notes 1, 2 and 13). SpaceX's own disclosure records purchasing $295 million of Megapack products in the same quarter, capitalised into PP&E. The $23 million difference reflects different measurement bases — Tesla's recognised revenue versus SpaceX's capitalised cost — and neither figure is wrong.

The incumbent mobile-satellite operators, on identical periods. Two comparisons are genuinely like-for-like because the quarters match exactly. For the three months ended 30 June 2026, Iridium Communications reported total revenue of $225.2 million (Iridium Communications Inc. Form 10-Q, quarter ended 30 June 2026) and Globalstar reported $64.8 million (Globalstar, Inc. Form 10-Q, quarter ended 30 June 2026). SpaceX's Connectivity segment did $4,291 million in the same three months — nineteen times Iridium and sixty-six times Globalstar. Globalstar's own filing is dominated by the pending Amazon mergers, listing among its risks "the ability of certain of our customers to terminate or amend contracts upon a change of control" and a merger-consideration adjustment tied to satellite milestones. The narrowband incumbents are not competitors to Starlink; they are spectrum inventory, and Amazon has bought one of them.

China. Guowang and Qianfan have launched roughly 190 and 162 satellites respectively as of mid-2026 on trade-press counts (Tier 3). Against Starlink's 9,600 that is not yet a competitive constraint. Against a stated Guowang ramp of 900 satellites in 2027 and 3,600 a year from 2028, it is a dated question rather than a settled one — and it is largely irrelevant to SpaceX's addressable market, since neither constellation will be licensed in the United States or most of Western Europe. Their real effect runs through launch demand and through third-country market access.

Part IIInvestment Brief — conclusion first

A structural assessment of economic position. No price targets, no entry or exit levels, no position sizing. This is not investment advice.

01  Where the rent is, stated plainly

The only structurally defended excess return in SpaceX today sits in Connectivity, and it rests on two things a rival cannot buy — an orbital shell at critical density and, as of May 2026, exclusive nationwide D2D spectrum — while the launch business earns a cost advantage it deliberately does not price, and the AI business earns a spread on rented demand it does not control.

The prevailing IPO narrative is that SpaceX is a launch monopoly with a broadband annuity attached and a moonshot on top. Every part of that is either wrong or incomplete. Launch is not a monopoly in the economic sense — it is a low-cost producer that wins about 60% of national-security missions at roughly two-thirds of its rival's average obligated value, and it charges its largest customer nothing at all, because its largest customer is itself. Broadband is an annuity, but its price fell 22% year over year while volume doubled. And the moonshot the equity is being asked to fund is not Mars; on the capital numbers it is a data-centre business.

The Space grade needs defending against this report's own criterion. The thesis says durable value requires a hard-to-replicate right, and the Space segment plainly lacks one — anyone may build a rocket, ULA and Blue Origin do, and reusability is diffusing. So why "neutral" rather than "exposed"? Because a lowest-cost producer earns a real, if smaller, return without any pricing power at all: the 65.8% gross margin on third-party work, struck on a cost base that did not move when volume fell 18%, is genuine surplus. It is competed-against rather than protected, and it will compress as Neutron and New Glenn certify. Neutral, not advantaged — and not exposed while the cost gap remains as wide as the FY2026 award values imply.

02  Connectivity: advantaged — but understand precisely why

The temptation is to credit the moat to the satellites. Nine thousand six hundred spacecraft is an impressive number and a poor explanation, because satellites depreciate and can be replaced; the shell turns over roughly every five years and each generation costs money. If the constellation were the moat, the moat would need rebuilding every five years, and a rival with $10 billion a year could eventually build one too. Amazon is trying exactly that.

The defensible answer has three parts, and only one of them involves hardware.

Density economics. Connectivity's cost of revenue fell from 52.0% to 48.0% of segment revenue while subscribers doubled and price fell 22%. DepreciationThe slice of a long-lived asset's cost charged against profit each period as it ages toward replacement. No cash moves when it is booked — the money went out when the satellite was built and launched — but it is a real cost, because the asset will have to be bought again. of $805 million spread across 11.15 million average subscribers is $24.07 per subscriber-month — against ARPU of $66. Every incremental subscriber on an existing beam is nearly pure margin. This is a network effect in the strict sense: the cost per user falls as users are added, so the operator who reaches density first can price below any entrant's cost and still earn. It is the same shape as cable, and it produced the same outcome.

Vertical integration of the deployment cost. Because launch is transferred at cost, Starlink's satellites sit on the books at a number no competitor buying commercial launch can match. Amazon's own 10-Q shows $7.4 billion of prepaid launch deposits — cash out of the door to third parties, including to SpaceX. The competitor is funding the incumbent's launch business in order to compete with the incumbent's broadband business. That is not a metaphor; it is a line in Amazon's balance sheet.

Spectrum. Until May 2026 this was the missing leg. Starlink was a broadband service requiring a terminal. With 65 MHz of exclusive nationwide mid-band, the addressable device becomes every phone. The FCC said so: it expects SpaceX "to be able to compete with incumbent terrestrial wireless providers with the beneficial result of potentially lowering prices for consumers," and drew the analogy to fixed wireless access, which "first slowly, then suddenly — revolutionized the market for in-home broadband."

If Starlink's advantage is density economics plus at-cost launch plus spectrum, then the cheaper Starship makes launch, the wider the Connectivity moat becomes — and the narrower the Space segment's. A vehicle that deploys 60 V3 satellites per flight at twenty times the capacity per launch collapses Starlink's cost per delivered gigabit. It also collapses the market price of third-party launch, because that same vehicle is available to customers. SpaceX's own success at Starship therefore transfers rent from the segment that sells launch to the segment that consumes it. The market is applauding Starship as a launch-business catalyst. On the structure of the accounts, it is a Connectivity catalyst that happens to run through the launch pad.

The best counter-argument against this verdict is the replacement cycle. Starlink satellites have a working life of roughly five years, so a 9,600-satellite shell implies replacing on the order of 2,000 spacecraft a year in perpetuity. The filings show the cost arriving: Connectivity depreciation and amortisation rose from $569 million to $805 million year over year, up 41.5%, and gross satellites on the balance sheet rose from $11,949 million at December to $13,788 million at June — $1,839 million added in six months. On that trajectory D&A grows faster than subscribers if the shell keeps expanding, and the margin expansion documented above reverses. The rebuttal is that V3 changes the arithmetic: one Starship flight replaces roughly twenty Falcon flights' worth of deployed capacity, so the replacement cost per delivered gigabit falls even as the satellite count rises. That rebuttal depends entirely on Starship, which is why the two verdicts are linked and why the Connectivity falsifier below is a margin test rather than a subscriber test.

Second order. If cheap heavy lift commoditises orbital capacity generally, the scarce input stops being the satellite and becomes the spectrum and the market access. That is the reading which explains both SpaceX's $19.6 billion EchoStar purchase and Amazon's $10.9 billion Globalstar purchase, three months apart, by two companies that own their own rockets. Two of the best-informed buyers in the industry independently concluded that the bottleneck had moved downstream to the licence. Competitors rarely agree that precisely.

03  Space: neutral — the price-setter problem

The mandate asks who sets the launch price when the largest customer is in-house. The filings answer it directly: nobody does. There is no internal price. There is a cost, capitalised.

First, no external observer can compute SpaceX's true launch margin, because 87% of its lift volume never passes through a revenue line. Second, the company faces no internal pressure to price launch efficiently — a Starlink programme manager buys rides at accounting cost, which is the cheapest possible transfer price and guarantees over-consumption relative to a market-priced world. Third, and most importantly for anyone valuing the Space segment: because the surplus is transferred rather than booked, the Space segment's reported economics understate the launch business and the Connectivity segment's overstate the broadband business, by the same amount. Any sum-of-the-partsValuing each division of a company on its own and adding the results, on the theory that the market is mispricing the bundle. It works only if each division's reported numbers mean what they appear to mean. that values the two separately at different multiples is arbitraging an accounting convention.

On third-party economics, the picture is a strong cost position weakening at the margin. The FY2026 national-security assignments give SpaceX five missions at an average obligated value of $143 million against ULA's $214 million across two. Phase 3 Lane 2 allocates roughly 60% of about 54 missions to SpaceX through FY2029. Being the cheap bidder that wins the most missions is a good business. It is not an annuity, because the mechanism producing it — reusable boosters — is being replicated. Rocket Lab is building Neutron. Blue Origin's New Glenn awaits certification, at which point up to seven Phase 3 missions are earmarked for it.

The bottleneck question: what happens to third-party launch rent if capacity doubles? SpaceX flew 77 Falcons in the first half. Add a working Starship at even modest cadence and the supply of kilograms to orbit rises by a multiple, not a percentage, while third-party demand — 17 customer launches in six months — is roughly flat. In any market where supply multiplies against flat demand, price falls to the marginal producer's cost. SpaceX will be the marginal producer, so it will capture volume; it will not capture price. That is the definition of a cost advantage without pricing power, and it is why the Space segment grades neutral rather than advantaged despite a genuine technological lead.

The alternative reading, and why it was rejected: national-security certification, Cape and Vandenberg pad access, and the FAA's 195-launch annual authorisation ceiling could all be read as regulatory barriers that will keep rivals out and preserve price. Certification is real but temporary — ULA already has it and Blue Origin is working through it. Pad access is a constraint on the incumbent as much as on entrants. The condition that would reopen this alternative is a durable, defence-driven bifurcation of the launch market in which national-security missions command a persistent premium that competitors cannot bid away — and the FY2026 award values point the other direction, with the incumbent as the cheap bidder.

04  Starship: the option, and what the equity is actually being asked to fund

Starship consumed $1,076 million of R&D in the June quarter, 1.12 times the entire Space segment's revenue, and over $15 billion cumulatively. What does that buy?

Three distinct claims, with very different evidentiary status. The Starlink V3 deployment claim is near-term, quantified and partially demonstrated — Flight 13 deployed 20 production V3 satellites, against a design target of 60. The Artemis lunar claim is contracted but contested; NASA has publicly opened the Artemis III lander to competition on the grounds that SpaceX is behind, and the propellant-transferPumping cryogenic fuel from one spacecraft to another in orbit — ship-to-ship refuelling. A Starship that reaches orbit has spent most of its fuel getting there, so anything beyond orbit depends on being topped up once it arrives. demonstration on which the whole architecture depends has not been performed. The point-to-point and Mars claims carry no contracted revenue in any filing opened in this session, and no verdict here rests on them.

A fourth use has quietly become the largest. If the orbital data-centre application is serious — up to a million satellites at roughly 100 kW each — then Starship's economic purpose is to lift compute, and the AI segment is the customer that justifies the vehicle. That closes the loop between the three segments in a way the IPO narrative never articulated: Connectivity's cash funds Starship; Starship lowers Connectivity's deployment cost and enables orbital compute; orbital compute is the AI segment's escape from the terrestrial power constraint that has the company burning gas turbines in Mississippi and litigating with the NAACP over it.

This is not a retrospective reading imposed on the company. The CFO framed it himself during the roadshow, in a transcript the company filed with the SEC: launch is "the start of it … it's hard to be a space company if you do not have assured access to space," and the interview then moves directly from Starship's mass-to-orbit economics to "gigawatts of orbital compute" and to the observation that SpaceX is "selling compute terrestrial to competitors" while the model remains "vertically integrated … but also open at each level" (Free Writing Prospectus — transcript of Gavin Baker's interview with CFO Bret Johnsen, filed 9 June 2026). Selling compute to your competitors is a coherent strategy for a capacity owner in a shortage. It is not a strategy that produces switching costs.

Whether one finds that visionary or alarming, Starship is a real option with a demonstrated technology path, an undemonstrated economic path, and a burn rate of roughly $4.3 billion a year. It is not a science project — 33 Raptors lit, a booster recovered, a relight achieved and 20 satellites deployed is engineering progress few organisations could match. It is also not yet an annuity, and it will not become one until a Starship upper stage flies twice.

05  AI: exposed — the ninety-day franchise

One number makes the case on its own. AI segment revenue rose from $818 million in March to $2,561 million in June, up $1,743 million. AI Segment Adjusted EBITDA rose from $(609) million to $1,146 million, up $1,755 million. The incremental Adjusted EBITDA margin on the new revenue is 100.7%.

No business has a 100% incremental margin. Adjusted EBITDA is profit before interest, tax, depreciation and amortisation, with a further list of items management judges unrepresentative added back on top of that. For most companies it is a rough stand-in for cash generation, and a serviceable one. For a company whose product is the use of machines it bought last quarter, it removes the cost of the thing being sold. The arithmetic is telling you what the cost is made of: the marginal cost of serving a reserved-capacity cloud contract is almost entirely depreciation, and Adjusted EBITDA adds depreciation back. AI segment D&A was $1,885 million in the quarter — 73.6% of AI segment revenue. Strip the add-back and the segment lost $1,257 million.

The nominal value of the Anthropic and Google agreements is roughly $75.4 billion. Neither party is obliged to pay it. Anthropic may exit on 90 days' notice after the initial three-month period; Google may exit on 90 days' notice after 31 December 2026, and may terminate immediately if SpaceX misses its 30 September 2026 GPU delivery. The company's own backlog definition captures this honestly, which is why total group backlog of $47.5 billion is smaller than the two contracts' nominal sum.

The other side of the ledger: $22.2 billion of non-cancellable obligations due in 2027 (Form 10-Q, Note 16 — Unconditional Obligations: exactly $22,244 million of $27,955 million total), $15.8 billion of capex in a single quarter, and $13.3 billion of related-party equipment debt at an implied all-in charge of at least 9.8%. The company has bought three years of assets against ninety days of revenue rights.

The counterparty concentration compounds it. Customer B was 19.5% of consolidated revenue — roughly $1,524 million — and relates solely to the AI segment. The 10-Q's newly added risk factor is unusually candid for a first quarterly filing: "a significant portion of our AI infrastructure revenue is concentrated in a small number of customers … the loss of a significant customer, the termination or non-renewal of one or more of these agreements … or shifts to internally developed infrastructure … could materially adversely affect our AI segment and consolidated revenue" (Form 10-Q, Item 1A — Risk Factors).

"Shifts to internally developed infrastructure" is doing a lot of work in that sentence. Google — which pays $920 million a month for SpaceX compute, sits on the SpaceX board through Donald Harrison, and is a party to the Investors' Rights Agreement — designs and operates its own acceleratorsThe specialised chips that do the arithmetic behind AI models — GPUs and the custom silicon built to replace them. "Accelerator" is the generic word for the part of a machine that handles the heavy lifting an ordinary processor cannot. at hyperscaleThe scale at which a handful of firms run computing: their own data centres, their own networks, increasingly their own chips. A hyperscaler renting capacity from somebody else is doing so by choice, not for want of the capability.. It is renting capacity because capacity is scarce this year, not because it lacks the capability. When the scarcity eases, so does the contract.

One second-order consequence decides the segment's grade. In a shortage, the merchant with capacity captures rent; the price is set by the buyer's willingness to pay rather than the seller's cost. But merchant compute has no switching cost — models and data belong to the customer, as both agreements expressly provide — and the asset has a life measured in a handful of years. So the rent is real and it is temporary, and it terminates on the earlier of the shortage ending or ninety days' notice. Meanwhile the depreciation runs for the full life of the asset regardless. The revenue has a 90-day duration and the cost has a five-year one. That mismatch is the whole risk, and no amount of demand growth fixes it.

06  The bottleneck test — five scarce nodes, and what happens if each doubles

07  Incentive map — who wants what, and where the rent lands as a result

08  Reading the quarter against $1.44 trillion

First, correct the denominator, because a great deal of commentary has it wrong. At $109.40 on 6 August 2026 against 13,181,779,945 shares outstanding, the equity value is $1.442 trillion (stockanalysis.com, SPCX quote, 6 Aug 2026; cross-check Yahoo Finance SPCX, 12:18 EDT; share count from the 10-Q cover page). Net of $100.0 billion of cash and securities and $38.4 billion of debt, enterprise value is $1.381 trillion. The two measures answer different questions. Equity value is what the shares cost. Enterprise value is what the whole business costs — the shares plus the debt a buyer inherits, less the cash a buyer gets to keep — and it is the number to use when comparing companies that finance themselves differently. Any figure in the $600–700 billion range in circulation is not reconcilable with the disclosed share count at any price the stock has traded.

Now the ratios, computed on the disclosed quarter.

The path to GAAPGenerally Accepted Accounting Principles — the American rulebook a listed company must follow in its audited accounts. GAAP profit is profit as the rules define it, before management adds back anything it considers unrepresentative. profitability is arithmetically visible and not near. Consolidated operating loss was $143 million. Connectivity earns $1,656 million; Space loses $542 million and AI $1,257 million. Hold Space's Starship spend flat and the group crosses into operating profit the moment AI's loss narrows by $143 million — which, on the Q1-to-Q2 trajectory, could happen next quarter. But that is operating profit, not free cash flow, and it is the wrong measure for this business.

Free cash flowThe cash left after a business has paid for the assets it must buy to keep running and growing — operating cash flow less capital expenditure. It is what is genuinely available to repay debt, buy things, or hand back to shareholders, and it is far harder to shape with accounting choices than profit is. is the honest test, and it is brutal. Operating cash flow for the six months was $3,466 million. Capital expenditure was $28,476 million. The gap is $(25,010) million in six months (Form 10-Q, MD&A "Summary of Cash Flows"; capex from Exhibit 99.1). At that rate the $100 billion of cash and securities funds roughly eight quarters. That is not distress — it is two full years of runway with an untouched $5 billion revolver and demonstrated access to $25 billion of investment-gradeA credit rating in the top tier, meaning the agencies judge default unlikely. The practical consequence matters more than the label: pension funds and insurers are largely confined to this category, so an investment-grade borrower can reach a far deeper pool of lenders, and at a lower rate. debt at 5.855%. It is, however, a company that has pre-committed its IPO proceeds.

Which reframes the offering. SpaceX raised $85.7 billion in equity and $25 billion in debt — $110.7 billion — and in the same six months spent $28.5 billion on capex and committed $28.0 billion more, of which $22.2 billion is due in 2027. The IPO was not an exit for a mature business. It was the funding event for a data-centre build.

09  A valuation framework, not a target

Two framings, deliberately different, because the honest answer to "is $1.44 trillion right?" is that it depends entirely on which business you think you are buying.

Framing one: what multiple is embedded? Take annualised Q2 revenue of $31.3 billion, grow it over the next twelve months, apply a range of enterprise-value multiples, add net cash of $61.6 billion, divide by 13.18 billion shares. The grid below brackets the price rather than arguing with it.

Framing two: what does the residual cost? This is the more useful question, and it is the one Grant would ask. Value Connectivity generously on its own annualised economics — revenue $17.2 billion, operating income $6.6 billion, Adjusted EBITDA $10.4 billion. Put it at 40 times annualised operating income, a rich multiple for any network business and 15.4 times its own revenue, and Connectivity is worth about $265 billion. Subtract that from the $1,381 billion enterprise value and the residual for Space plus AI is $1,116 billion.

Those two segments together produced $3,523 million of revenue in the quarter — $14.1 billion annualised — and lost $1,799 million at the operating line. So the market is paying roughly 79 times revenue for two loss-making businesses, one of which is a launch operation with declining cadence and the other a compute rental on 90-day contracts. Even at 60 times Connectivity's operating income the residual is $983 billion, or 70 times.

10  The lock-up expiry and share overhang

Two facts decide this section, and both were fixed by contract before the stock ever traded: 911.5 million shares came free on 6 August 2026 — the evidence-cutoff day — and the price condition that would have released a further 455.8 million alongside them failed by 40%.

The schedule below is reconstructed tranche by tranche from the prospectus. It is not an estimate; the prospectus prints it.

The float arithmetic. The float is the part of a company that can actually change hands — shares outstanding, less everything locked up, held tight by insiders, or otherwise off the market. It matters more than the share count, because price is set by the shares available to trade, not by the ones sitting still. Before 6 August, roughly 691 million shares traded freely — about 5.2% of the 13.18 billion outstanding, derived by subtracting the two locked pools (≈7,800 million long-dated plus ≈4,691 million in the 180-day pool) from shares outstanding. Adding 911.5 million takes the float to roughly 1,602 million: a 2.32× increase in a single day. By 8 December, if every tranche releases as scheduled, the tradeable float rises to roughly 5.4 billion shares — an eight-fold expansion from where price discovery has been happening.

The absorption arithmetic. Average daily volume over the thirteen sessions from 20 July to 5 August was 84.5 million shares; excluding the two earnings-driven sessions it was 67.9 million. So 911.5 million eligible shares represent between 10.8 and 13.4 days of normal trading. That is the right way to size it — and the wrong way to conclude from it, because eligibility is not supply.

Most published lock-up analysis goes wrong here, because the incentives matter more than the arithmetic. Of the 911.5 million shares, none belong to Musk, and none belong to the extended-lock-up holders who agreed to wait until 2027. The pool is employees and early non-affiliate investors. Their incentives diverge sharply:

• Employees hold shares acquired through option exercises and ESPP purchases — the 10-Q discloses 8.5 million option-exercise shares and 3.2 million ESPPEmployee Stock Purchase Plan — a scheme letting staff buy company shares through payroll deductions, usually at a discount to the market price. shares issued between 1 April and 12 June alone, plus 25.9 million RSUsRestricted Stock Units — a promise of shares that vest over time or on meeting a condition. Until they vest they are not shares at all; once they do, the employee owns stock and may sell it, which is the pipe through which employee pay becomes market supply. granted. Many have held for a decade through repeated tender offers at fair market value. Their marginal seller behaviour is diversification, not conviction — and it is highly price-elastic in the wrong direction: a stock below its IPO price makes employees who bought the story less likely to sell, not more.
• Early venture holders face the opposite calculus. The Valor Entities' Rule 10b5-1 plan, adopted on the first day of trading, contemplates distributing up to 225,857,490 shares to limited and general partners through 30 September 2027. A distribution is not a sale — it moves stock to LPs who then decide individually — but it is the mechanism by which a fifteen-year-old venture position becomes public-market supply, and it covers 44.9% of Valor's 503.4 million shares.
• Officers have pre-committed the other way. The CFO's plan "does not commence sales until 2027"; the President and COO's plan runs to 30 June 2027; both "agreed to subject the vast majority of [their] shares to the extended lock-up period." Combined coverage is 1.5 million shares — 0.01% of the company. On the disclosed evidence, insider selling by named executives is not the overhang.

The historical analogue, used carefully. The standard reference point for staggered mega-cap unlocks is that the price impact is front-loaded into the announcement window rather than the release date, because the schedule is disclosed in the prospectus and therefore knowable months ahead. SPCX's behaviour is consistent with that: the stock fell from $225.64 to the $110s over the seven weeks before 6 August, and traded up 1.04% on the morning of the release itself. The efficient-markets reading is that the unlock was priced in June. The uncomfortable counter-reading is that it cannot have been fully priced, because until 4 August nobody knew whether the 455.8-million-share price tranche would trigger — and that uncertainty resolved adversely only forty-eight hours before the release.

One filing made on the day of the print sizes part of the employee supply precisely. Alongside the 10-Q, SpaceX filed a Form S-8The short registration form a company uses for shares issued to its own staff under equity plans. The reoffer prospectus attached to it lets those employees resell the shares into the market — it makes selling legally possible, and says nothing whatever about whether anyone intends to. containing a reoffer prospectus registering 136,949,657 shares of Class A common stock for resale by "current and former employees and other service providers" who acquired them through RSU settlements and option exercises under nine separate equity plans — SpaceX's 2015, 2024 and two ESPP plans, plus the Swarm, xAI 2023, xAI 2025, X 2023 and Mesh Optical plans (Form S-8 reoffer prospectus, filed 4 August 2026). Two features matter. It is expressly "subject to any lock-up and market standoff agreements applicable to such Selling Stockholder," so registration does not release anyone. And each seller is capped at the Rule 144(e) volume limit in any three-month period — for an affiliateIn securities law, an insider close enough to the company — an officer, a director, a controlling holder — that their sales are capped and reported. An ordinary employee is not an affiliate; a founder with 84.4% of the votes plainly is., broadly 1% of shares outstanding or the average weekly volume, which at SPCX's size is not a binding constraint for any individual employee but is a structural brake on concerted selling. Registering 136.9 million shares two days before the unlock is the company doing exactly what an issuer does when it expects employees to sell: making the mechanics orderly.

What the evidence does not support. No filing discloses how many of the 911.5 million shares are actually intended for sale. No announced secondary offering exists in any document opened this session. The registration rightsA contractual right, negotiated long before a company goes public, to compel it to register your shares for public sale at its own expense — so that a large early holder can exit through a formal offering rather than dribbling stock into the market for months. under the Investors' Rights Agreement — held by entities affiliated with Musk, Google, Valor and DFJ Growth — are exercisable by holders of a majority of registrable securities beginning six months after effectiveness, which is 11 December 2026, and none has been exercised. Goldman Sachs, as representative, retains sole discretion to release any shares early; no such waiver has been disclosed. Anyone quoting a precise expected selling volume is inventing it.

The strongest counter-argument, and it is a good one. SPCX entered the Nasdaq-100 Index before the open on 7 July 2026 — fifteen trading days after listing, under fast-entry rules that let the index temporarily exceed 100 constituents rather than removing anyone. Estimates at the time put mandatory Nasdaq-100 buying at roughly $4.3 billion, with about $3 billion more from Russell reweighting (Nasdaq-100 inclusion timing circulated in trade press in late June 2026 — Tier 3 / not independently verified this session; the underlying Nasdaq index press release was not opened, and the dollar estimates are third-party, not official).

It cuts against the simple overhang story. Major index weights are float-adjusted. If the tradeable float rises from roughly 691 million to roughly 5.4 billion shares between August and December — an increase of about 7.8 times — then every index fund tracking SPCX must buy more shares at each float-adjustment date, mechanically and without regard to price. The unlock schedule is therefore not purely a supply event. It is a supply event that automatically generates a slice of its own demand, and the slice grows with each tranche. This is the single best reason to expect the staircase to be absorbed better than the day-count arithmetic implies, and it is why 6 August traded up rather than down.

It is not a complete answer. Float-adjusted index buying scales with the index's assets and SPCX's sub-1% weight, not with the size of the unlock; on the estimates above the entire initial inclusion absorbed roughly $4.3 billion against a 6 August tranche worth about $98.7 billion at the prior close. Passive demand smooths the path. It does not clear the inventory.

The consequence chain. First: because the price test failed, 469 million additional shares now land on 8 December instead of 6 August, moving supply from a date with an earnings catalyst and saturation coverage to one with neither. Second: the sequencing means the float roughly quadruples before the founder's 6.4-billion-share release on 12 June 2027, which is either good design — the market learns to absorb stock in stages before the big one — or a slow-motion demonstration of how much stock is waiting. Third, and least discussed: the price at which the float expands sets the cost of two pending acquisitions, because the Cursor consideration is a fixed dollar value settled at a trailing VWAPVolume-weighted average price — the average price a share changed hands at over a stated period, weighted by how many shares traded at each level. Deals settle on it because it is much harder to push around than a single closing print. while the EchoStar equity leg is a fixed share count at $42.40 — which is the subject of the next section.

11  The dilution ledger — three commitments that grow as the stock falls

Three announced transactions will be settled in Class A shares, and two of them get more expensive in share terms as the price declines.

A merger settled in a fixed dollar amount of stock at a trailing VWAP transfers all the price risk to the acquirer's existing shareholders. If SPCX halves again before closing, the Cursor sellers receive twice the shares. The company retained no collarA ceiling and a floor written into a share-settled merger, limiting how far the number of shares issued can move if the buyer's stock swings before closing. Without one, the seller is indifferent to the price and the buyer's existing shareholders absorb every point of the move., no cap and no walkaway right disclosed in the 8-K — and it had already given up the ability to walk, since terminating the option would have triggered a $1.5 billion fee plus an $8.5 billion deferred services fee.

12  Governance, key man and the related-party web

The governance facts are not ambiguous, and each cuts in a specific direction.

Control is absolute and undated. 84.4% of votes, a ten-to-one dual class, no sunset in the certificate, no dividend, and a founder locked until 12 June 2027 with no early release. Outside shareholders have no mechanism to influence capital allocation. Whether that is a feature depends entirely on whether one thinks pouring Connectivity's cash into 86%-of-capex AI infrastructure is a good idea, and the market's answer on 5 August was a 13.6% decline.

Key-man exposure is unusually literal. The same individual is CEO of SpaceX and of Tesla, which holds a $3.0 billion stake in SpaceX and sells it $318 million a quarter of Megapacks; he acquired Twitter personally in 2022 and now sits atop the entity that owns it; and he is a named defendant alongside the company in the European Commission's DSAThe European Union's Digital Services Act, which sets obligations for large online platforms on illegal content, transparency and risk assessment — and lets the Commission fine them, at group scale, for failing to meet those obligations. decision, which imposed a €120 million fine on 5 December 2025 on "XIUC, X., x.AI, and Elon Musk (together, the 'parties')," now under challenge in the General Court (Form 10-Q, Note 16 — legal proceedings).

The related-party web has a cost that can be estimated. $13.3 billion of AI equipment financing with a director's firm at an implied all-in charge of at least 9.8% annualised, while the company borrows publicly at 5.855% and sits on $100 billion of cash. Four percentage points on $13.3 billion is roughly $530 million a year of avoidable expense if a cheaper structure were available. The filings do not say why the structure was chosen and there may be good reasons; the company simply does not offer one.

Shareholder remedies were narrowed on the way public, and nobody remarked on it. SpaceX is a Texas corporation, and its amended and restated bylaws provide that "the sole and exclusive forum and venue for Internal Disputes … shall be the Texas Business Court, Eleventh Division," together with a capitalised jury waiver for such disputes (Amended and Restated Bylaws, Exhibit 3.2 to the Form 10-Q, § 10.1). Combine an 84.4% voting founder, a board whose non-employee members are unpaid, and internal disputes routed to a newly created specialist court without a jury, and the practical avenue for a minority shareholder to challenge capital allocation is narrow. The company also listed on two venues — the Nasdaq Global Select Market and Nasdaq Texas, LLC — registering the Class A shares on both (Form 8-A12B, 10 June 2026).

The litigation tail is qualitatively different from a normal industrial's. Beyond the DSA fine and a $172 million patent judgment against Twitter under appeal, the company faces multiple putative class actionsLawsuits brought on behalf of a whole group of people claiming similar harm. "Putative" means the court has not yet agreed the group may sue as one — certification is a separate fight, and it is the fight that decides the size of the exposure. arising from Grok's image-generation features alleging creation of non-consensual sexualised imagery, including of children, plus an action by the Mayor and City Council of Baltimore. Total accrued litigation loss is $354 million. These are xAI-inherited risks that a launch and satellite company would never have carried, and they now sit inside the same reporting entity that holds NASA and Space Force contracts.

13  SteelmanningBuilding the strongest possible version of the argument against your own — the opposite of a straw man. If a conclusion survives the other side stated at its best, it is worth considerably more than one that only survives the other side stated at its worst. the bull case — then taking it apart

The best version of the opposing view is not "space is cool." It is a serious argument, and it goes like this.

The strongest single plank is that at $100 the market prices the AI business at nothing. Figure 24 shows the opposite. Value Connectivity at a generous 40× annualised operating income and the residual for Space plus AI is $1,116 billion on $14.1 billion of annualised, loss-making revenue. The market is not pricing AI at zero; it is pricing Space and AI together at 79 times revenue. That plank does not survive contact with the segment disclosure the bull case did not have when the target was set on 24 July — eleven days before the first segment print.

The second plank, the power wall, is the best argument in the bull case and the one most worth holding. If terrestrial data centres genuinely run out of interconnect capacity, then a company with a reusable heavy-lift vehicle and a licensed constellation has an option no one else holds. But note what the company is doing while waiting: burning mobile gas turbines in Mississippi under a Clean Air Act challenge, and buying $295 million of Tesla Megapacks a quarter. Those are the actions of a firm solving the power problem on the ground, on the ground's timeline. The orbital answer is a filing, not a business.

14  Second- and third-order effects

The competitor funds the incumbent's rocket. Amazon carries $7.4 billion of satellite launch prepayments and has bought launches from SpaceX to close its FCC milestone gap. Every dollar Amazon spends attacking Starlink's market strengthens the launch business whose at-cost transfer is Starlink's advantage. The third-order effect: this is why SpaceX has no incentive to refuse Amazon capacity, and why launch pricing to rivals will stay competitive — the launch business is more valuable to SpaceX as an instrument of the Connectivity moat than as a profit centre.

Cheap launch destroys launch's rent and creates spectrum's. Walked through in Section 02. The generalisation is the useful part: when a bottleneck is relieved, the rent moves one step down the chain to whatever is still scarce. In 2015 the scarce thing was reaching orbit. In 2026 it is the licence to use the airwaves once you are there. In 2030, if a million-satellite orbital compute constellation is real, it will be the power and the ground-station spectrum to move exabytes down.

The lock-up staircase transmits SPCX's price into Tesla's earnings. Tesla marked its $2.0 billion SpaceX position to $3,007 million and booked a $1.00 billion gain in the June quarter, with sale restrictions expiring December 2026. A material decline in SPCX therefore produces a mark-to-market loss in a separate S&P 500 company's non-operating income, in a quarter that also contains the December lock-up releases. Two listed securities are now mechanically coupled through one man's cap table.

The bankruptcy nobody connected. Hughes Satellite Systems Corporation filed Chapter 11 on 2 August 2026. SpaceX filed its 10-Q on 4 August with no mention of it anywhere in the spectrum disclosure. The steelman for that omission is strong: HSSC is a subsidiary, EchoStar Corporation — the actual counterparty — did not file, and critically the licences already sit in Spectrum Business Trust 2025-1, transferred on 22 May 2026, which is precisely what a bankruptcy-remoteA structure that parks an asset in a separate legal entity so that if the seller later collapses, the asset cannot be pulled back into the seller's bankruptcy estate. Lawyers build them for exactly the situation that arrived here. two-step structure is for. SpaceX and its counsel appear to have structured around this exact risk a year before it materialised, which is a genuine credit to the deal team. But two things remain open: $856 million of forgivable "loans" already advanced to the trust are prepaid consideration with no closing, and up to $8.5 billion of the purchase price is a payoff of "designated EchoStar debt" whose priority and amount a bankruptcy court now influences. The absence of any reference in a filing made two days later is a disclosure choice worth noticing.

15  Four non-consensus conclusions

16  Falsifiers, catalysts and dated predictions

A closing thought, offered as perspective rather than conclusion. Investors have been here before with companies that were simultaneously extraordinary and expensive, and the lesson from those episodes was rarely that the business was a mirage. It was that the price already contained the future, and that the future arrives slowly enough to test the patience of anyone who paid for it in advance. SpaceX has built something no institution — governmental or private — had managed in sixty years of trying. Whether it has built it for the shareholders who bought it in June, or for the ones who will buy it from them at some later date and some other price, is a different question, and the first quarterly report does not settle it.

Appendix AEvidence register

Every source below was opened and read in the course of preparing this report. Tier 1 = issuer filings with the SEC, regulator orders and agency publications. Tier 2 = reputable trade or wire reporting. Tier 3 = single-source, aggregated or estimated material, used only where flagged and never load-bearing. Distinct sections of a single filing are listed once with the sections named. Evidence cutoff: 6 August 2026.

Appendix BNot verified / not load-bearing

The following could not be confirmed from a source opened in this session. Each is recorded here explicitly, and none carries a conclusion, a grade or a confidence level anywhere in this report.

1. The identity of Customer A and Customer B. The 10-Q discloses that Customer A was 18.3% of consolidated revenue across all three segments and Customer B 19.5% within AI, and names neither. The report's suggestion that Customer A is likely the US government in aggregate and Customer B likely Anthropic is labelled as inference in the text. No verdict depends on either identification.
2. Current status of the Artemis III lander re-competition. The October 2025 SpaceNews report of NASA opening the contract is the most recent primary-adjacent account opened. No NASA, GAO or OIG document dated 2026 was retrieved. The Starship verdict is stated as an option with undemonstrated economics and does not rest on any assumed Artemis outcome.
3. Any current SpaceX cost per kilogram to orbit. The filings cite NASA estimates for the 2010 Falcon 9 ($2,700/kg) and the 2018 Falcon Heavy ($1,400/kg) and a forward aspiration of a 99% reduction against an $18,500 historical average. No current figure is disclosed anywhere and none is asserted here.
4. The specific parameters of the orbital data-centre FCC application. The Space Bureau's acceptance-for-filing notice was blocked to automated retrieval and no FCC primary was opened. The "up to a million satellites" characterisation rests on the SpaceNews account of the 22 July 2026 order, which describes those applications as excluded from the new streamlined timelines. The report treats orbital compute as a stated ambition, not a business.
5. The FCC's July 2026 Upper C-band decision regarding SpaceX's satellite request. The relevant order and the trade account of it were not opened (one returned HTTP 403). The passage in Section 09 is flagged Tier 3 in place and carries no conclusion.
6. Eutelsat / OneWeb financials. No Eutelsat primary filing or results release was opened; one fetch returned HTTP 429. The €280m LEO revenue figure is trade-press and is used only to establish an order of magnitude in the competitive section. The competitive verdict rests on Amazon's FCC order and Amazon's own 10-Q instead.
7. Chinese constellation satellite counts. Guowang at roughly 190 and Qianfan at roughly 162 as of mid-2026 are trade-press aggregations. No Chinese regulatory or operator primary was opened. The report's conclusion — that neither is a near-term constraint on SpaceX's addressable market — would hold across any plausible value of these counts.
8. The claim that roughly 260 Starlink satellites were decommissioned in the first half of 2026. Encountered in a falsification search and not opened or verified. The replacement-cycle argument in Part II §02 is instead built from the filings' own disclosed depreciation ($569m → $805m) and gross satellite balances ($11,949m → $13,788m).
9. Analyst consensus ratings and median price target. Aggregated figures were encountered in search results but the aggregating pages were not opened. Only the Morgan Stanley view, from a page that was opened, is cited.
10. Whether any "designated EchoStar debt" under the Spectrum License Purchase Agreement sits at Hughes Satellite Systems Corporation or its filing subsidiaries. No document identifies the designated debt by issuer. The Chapter 11 discussion is therefore framed as an open question about disclosure, not as a valuation adjustment.
11. The precise construction of the EchoStar equity consideration. The filing says "approximately $11.1 billion in equity, payable through the issuance of approximately 261.8 million shares … at a fixed value of $42.40 per share." This report reads the share count as fixed. A value-fixed reading is grammatically possible; no closing has occurred to settle it, and the alternative is flagged in Figure 27's method note.
12. The Nasdaq index release itself. The official Nasdaq investor-relations release timed out on retrieval. The 7 July 2026 effective date and the fast-entry mechanism rest on a Tier 3 secondary source that was opened, and the index-absorption argument is presented as a counter-argument with its own limitation stated.
13. Actual selling intentions behind the 6 August release. No filing discloses how many of the 911.5 million eligible shares are intended for sale; no secondary offering has been announced; no registration-rights demand has been made; and no underwriter waiver has been disclosed. Every statement about absorption in Part II §10 is arithmetic about eligibility and volume, never a forecast of supply.

This is an internal structural assessment of a company's economics and competitive position. It is not investment advice, not a recommendation, and not a solicitation. It contains no price targets, entry or exit levels, or position sizing. All figures are as disclosed in the sources listed in Appendix A; derived figures state their arithmetic in the accompanying method notes. Evidence cutoff: 6 August 2026. Market price and capitalisation are intraday as of 12:18 EDT on that date and will have moved.

Nasdaq-100 inclusion timing and rebalancing dollar estimates. Circulated in trade press (late June 2026). The specific Seeking Alpha URL cited in the draft returned HTTP 403 this session; the underlying Nasdaq index release was not opened. No conclusion, grade, or valuation step depends on index-inclusion flows.

Golden Dome line-item revenue. Program participation is supported by defense-press accounts of OTA prototype pools and by SpaceX's own aggregate Starshield award disclosure. SpaceX does not report a "Golden Dome" revenue line; no verdict depends on a branded programme total.